Machine learning-based transformer service quality management method and system
By combining Xgboost and Logistic models in a machine learning approach, a service quality management system for power distribution areas was constructed. This system addresses the issue of service quality evaluation from the customer's perspective, enabling intelligent analysis and personalized service decision-making, thereby improving the scientific nature of power supply service quality and customer satisfaction.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- 国家电网有限公司客户服务中心
- Filing Date
- 2023-01-05
- Publication Date
- 2026-06-05
AI Technical Summary
Existing power supply companies struggle to evaluate service quality from the customer's perspective in their distribution area service quality management. They lack intelligent analysis and early warning tools, fail to create customer profiles, lack personalized service decision support, and the evaluation of power supply service quality is significantly influenced by human factors.
By employing machine learning-based methods and combining Xgboost and Logistic models, and through data augmentation, customer behavior profiling, and entropy algorithms, a service quality management system for distribution areas is constructed to achieve customer satisfaction prediction, user profile identification, and service quality evaluation.
It has improved the intelligence level of service quality management in the distribution area, provided personalized service decision support, reduced human interference, and enhanced the scientific nature of power supply service quality and customer satisfaction.
Smart Images
Figure CN116433064B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of power big data mining, specifically to a method and system for managing the service quality of distribution transformer areas based on machine learning. Background Technology
[0002] With societal development, the continuous reform of power companies, and changes in market supply and demand, power supply companies have come to a profound understanding that the quality of power supply services directly impacts their survival and development. Therefore, improving the quality of power supply services is essential for the steady development of power supply companies, and only through scientific and reasonable evaluation of power supply service quality can problems be identified and addressed to improve service quality.
[0003] Analysis of historical work orders from the 95598 service hotline reveals that while power supply companies at all levels have invested significant resources in power grid construction, improvements in customer experience have been minimal. The existing service quality control model for distribution transformer areas suffers from three main problems: First, it is difficult to conduct distribution transformer area management quality evaluation from the customer's perspective. Second, there is a lack of intelligent analysis and early warning tools when implementing service risk assessments for distribution transformer areas. Third, there is insufficient cross-disciplinary coordination capabilities, resulting in a failure to create customer profiles and a lack of decision-making support tools for providing personalized services. Summary of the Invention
[0004] In view of this, the present invention proposes a machine learning-based method and system for managing the service quality of distribution transformer areas. It combines Xgboost and Logistics to construct new features for effective prediction of customer satisfaction; it constructs user profiles from the perspective of payment to identify high-value users; and it uses an entropy algorithm to evaluate the quality of power supply services, thereby comprehensively managing the service quality of distribution transformer areas.
[0005] The method of this invention is implemented using the following technical solution: a machine learning-based method for managing service quality in transformer substations, comprising the following steps:
[0006] The power work order data was analyzed and processed to obtain the feature variable dataset D;
[0007] Based on the principle of minimizing neighborhood risk, a data simulation neighborhood interpolation algorithm is constructed to achieve data augmentation and obtain an augmented dataset.
[0008] Based on the enhanced dataset, new features are obtained using the Xgboost model; the new features are then recombined with the feature variable dataset, and the recombined data is input into a logistic regression classifier to obtain the prediction results of customer business satisfaction.
[0009] Establish customer behavior profiles and perform cluster analysis on customers from the perspective of value segmentation;
[0010] Calculate the values of each indicator used for comprehensive evaluation in each business segment, construct an entropy algorithm to obtain the weight of each indicator, and calculate the service quality score of each distribution area.
[0011] Based on the combined results of customer business satisfaction forecasts, customer cluster analysis, and service quality scores for each service area, the service quality of the corresponding service areas is adjusted.
[0012] Preferably, the data augmentation process includes:
[0013] Form matrix X from the n d-dimensional data points of the feature variable dataset D;
[0014] Perform zero-mean processing on each column of matrix X;
[0015] Find the covariance matrix of matrix X;
[0016] Find the eigenvalues and corresponding eigenvectors of the covariance matrix;
[0017] Arrange the eigenvectors from top to bottom according to the size of their corresponding eigenvalues, and take the first b rows to form matrix P;
[0018] Dimensionality reduction of matrix X is performed based on matrix P, reducing matrix X to b dimensions, resulting in the dimensionality-reduced feature variable dataset D. p ;
[0019] For the dimensionality-reduced feature variable dataset D p For each data point d, calculate the distance between other samples belonging to the same category and data point d, and find the b samples that are closest to data point d, which are then considered as the b nearest neighbors of data point d.
[0020] Randomly select C samples from b neighboring samples and calculate the data augmentation samples;
[0021] Using iPCA, the inverse operation of principal component analysis (PCA), the data augmented samples are restored to their original dimensions.
[0022] Put the data from each original dimension into the augmented sample dataset, and then put the augmented sample dataset into the feature variable dataset D to obtain the augmented dataset.
[0023] Preferably, new features are obtained using the Xgboost model; the new features are then recombined with the feature variable dataset, and the recombined data is input into a logistic regression classifier to obtain the prediction results of customer business satisfaction; including:
[0024] Configure XGBoost model parameters, including the number of iterations and the training step size;
[0025] The enhanced dataset is split into a training set and a test set. The training set is used as the input to the Xgboost model. The feature vectors generated by the Xgboost model are processed by one-hot encoding to generate new features.
[0026] The new features are recombined with the feature variable dataset, and then input into a logistic regression classifier. The output of the logistic regression classifier is the prediction result of customer business satisfaction.
[0027] Preferably, customer behavior profiles are established, and customer clustering analysis is performed from the perspective of value stratification, including:
[0028] The RFM indicator was constructed using a sample of customer payment data, and a dataset was built based on the RFM indicator.
[0029] Using a dataset constructed based on the RFM index as input, a K-means clustering model is built, and the silhouette coefficient method is used iteratively to select an appropriate cluster K.
[0030] Randomly select N data objects as the initial cluster centers;
[0031] Calculate the distance of each remaining data object to the N initial cluster centers, and assign the data object to the cluster class containing the nearest cluster center;
[0032] Adjust the new class and recalculate the center of the new class.
[0033] Preferably, the indicators used for comprehensive evaluation include: data collection coverage, data collection success rate, line loss rate, theoretical line loss, electricity meter ratio, power outage duration, number of power outage records, power outage duration, number of power outage records, and voltage anomaly duration.
[0034] More preferably, the values of each indicator used for comprehensive evaluation in each service segment are calculated, an entropy algorithm is constructed to obtain the weight of each indicator, and the service quality score of each distribution area is calculated; including:
[0035] Based on the principle of entropy method, a mathematical model for the index system is established;
[0036] Based on the indicators, positive and negative values are selected for dimensionless data processing;
[0037] Calculate the weight of a particular scheme relative to that indicator;
[0038] Calculate the information quotient and redundancy of the indicators;
[0039] Calculate the indicator weights;
[0040] The service quality score for each district is obtained by weighting and summing the indicator weights and the standardized weight values.
[0041] The system of this invention is implemented using the following technical solution: a machine learning-based service quality management system for distribution stations, comprising the following modules:
[0042] The dataset acquisition module is used to analyze and process power work order data to obtain the feature variable dataset D;
[0043] The data augmentation module is used to construct a data simulation neighborhood interpolation algorithm based on the principle of minimizing neighborhood risk to achieve data augmentation and obtain an augmented dataset.
[0044] The combined model prediction module is used to obtain new features from the enhanced dataset using the Xgboost model; the new features are then recombined with the feature variable dataset, and the recombined data is input into the logistic regression classifier to obtain the prediction results of customer business satisfaction.
[0045] The customer clustering analysis module is used to build customer behavior profiles and perform clustering analysis on customers from the perspective of value stratification.
[0046] The service quality score calculation module is used to calculate the values of various indicators used for comprehensive evaluation in various services, construct an entropy algorithm to obtain the weight of each indicator, and calculate the service quality score of each station area.
[0047] The integrated module is used to combine the predicted results of customer business satisfaction, the results of customer cluster analysis, and the service quality scores of each service area to adjust the service quality of the corresponding service area.
[0048] Compared with the prior art, the present invention has the following advantages and effects:
[0049] 1. Existing technologies such as XGBoost and Logistic models are two commonly used algorithms with good prediction accuracy; however, they also have some shortcomings: XGBoost performs well in handling variables with interactions, but struggles to handle linear relationships between variables; similarly, Logistic can handle linear relationships between variables, but cannot truly handle interactions. Therefore, this invention combines the two models to form an XGBoost and Logistic combined model for effective prediction of customer satisfaction. XGBoost identifies feature combinations with interactions, constructs new features, and adds them to the original features for training the Logistic model, thereby maximizing the advantages of both models.
[0050] 2. From the perspective of business data, the amount of data showing customer dissatisfaction with electricity services is far lower than the amount of data showing satisfaction, which leads to an imbalance in the overall sample dataset. This invention uses data augmentation to increase the amount of dissatisfaction data, thereby solving the problem caused by data imbalance.
[0051] 3. This invention constructs user profiles from the perspective of payment to assist in refined operations. It uses the RFM+Kmeans model to stratify user value, identify high-value users, and achieve precise marketing.
[0052] 4. The power supply service quality evaluation in this invention adopts the entropy algorithm, which can more profoundly reflect the distinguishing ability of indicators compared with other algorithms, and thus determine the weights. By assigning weights to the indicators, this algorithm avoids the interference of human factors in the weights of each evaluation indicator, making the evaluation results more in line with the actual situation, and solving the problem that the weighting process of indicators in existing evaluation methods is greatly affected by human factors. Attached Figure Description
[0053] Figure 1 This is a flowchart of a machine learning-based service quality management method for distribution centers in an embodiment of the present invention. Detailed Implementation
[0054] This invention relies on data from the State Grid Corporation of China and comprehensively utilizes big data mining technology to analyze and study the selection of various indicators and model design for external and internal evaluations of power supply service quality. The selection of indicators largely employs customer quality perception and service blueprint theory, while the service quality model design utilizes methods from data mining, such as statistics and machine learning. This research on comprehensive service quality evaluation methods lays the foundation for designing a scientific and reasonable comprehensive evaluation system for power supply service quality, and also serves as a reference for research on power supply service quality evaluation in different regions. Then, to verify the operability and practicality of the designed evaluation system, a power supply company is used as an example. After analyzing the current situation of the power supply company, the aforementioned service quality evaluation design method is applied, combined with the actual situation of the power supply company, to establish a service quality evaluation system suitable for that company. Through the evaluation analysis of the established evaluation system, deficiencies in the service quality of the power supply company are identified, and a series of improvement measures and suggestions are proposed. The evaluation process and results of this power supply company systematically illustrate how to establish a scientific comprehensive evaluation system based on reality, and also confirm the practicality and effectiveness of the evaluation system.
[0055] This invention establishes three core models in the construction of the entire district service quality evaluation system: a district service quality evaluation model, a customer business satisfaction model, and a customer group profiling model. The output of the customer business satisfaction model can serve as some feature variables in the district service quality evaluation model. The entire district service quality management system will be developed more from the customer's perception perspective.
[0056] The present invention will be further described in detail below with reference to the embodiments and accompanying drawings, but the embodiments of the present invention are not limited thereto.
[0057] Example 1
[0058] This embodiment provides a machine learning-based method for managing the service quality of power distribution areas. First, it establishes a combined XGBoost and Logistic model, along with the DSNI (Data Simulation Neighborhood Interpolation) algorithm. Using residential customers as the core target group, it analyzes massive call order data from the 95598 hotline to predict customer satisfaction. Then, based on customer order data, it establishes customer profiles along the payment dimension, constructs a three-dimensional dataset using RFM indicators based on payment data, and utilizes k-means clustering for customer value stratification. Finally, it designs a power supply service quality evaluation system, including external and internal evaluations. By establishing an internal and external service quality evaluation index system, it calculates the power supply service quality evaluation score by index level and business area, comprehensively achieving the management of power distribution area service quality.
[0059] The service quality management method for the transformer substation in this embodiment, such as Figure 1 As shown, the specific steps include:
[0060] S1. Analyze and process the massive amount of power work order data from 95598 to obtain a dataset of characteristic variables; the data sources include complaint work order data from power companies in various provinces and call work order data from the State Grid customer service center.
[0061] This step preprocesses the massive amount of work order data, specifically including:
[0062] 1) Remove work order data records with outliers or missing values;
[0063] 2) Selecting and constructing feature variables: For example, select feature indicators such as customer call time, call duration, number of reminders, and business type. The feature variable dataset D extracted after processing is shown in Table 1.
[0064] Table 1
[0065]
[0066] S2. Based on the VRM (Vicinal Risk Minimization) principle, the DSNI algorithm is constructed to achieve data augmentation, obtain an augmented dataset, and solve the problem of imbalanced data samples.
[0067] This step is based on the principle of neighborhood risk minimization. It uses the DSNI algorithm to construct a fusion algorithm of Xgboost and Logistic Regression, DSNI_Xgb_LR, to augment the feature variable dataset extracted in step S1 and construct more samples based on the original samples.
[0068] In this embodiment, this step involves determining the number of neighboring samples b and the amount of data augmentation C for each sample in the feature variable dataset D (where the feature flag represents unsatisfactory samples) after preprocessing in step S1, in order to achieve data augmentation and resolve the sample imbalance problem. The value of C is less than b; in this embodiment, b is chosen to be 12 and C to be 6. The specific steps are as follows:
[0069] 1) The n (1 million data points in this example) and d-dimensional (12-dimensional data points in this example) data points of the feature variable dataset D are used to form a matrix X. A sample of the data in matrix X is shown in Table 2.
[0070] Table 2
[0071]
[0072] 2) Perform zero-mean normalization on each column of matrix X (representing a field, such as call zone, call time, etc.), that is, subtract the mean of that column. The formula for normalization is as follows:
[0073]
[0074] 3) Calculate the covariance matrix of matrix X using the following formula:
[0075]
[0076] 4) Find the eigenvalues and corresponding eigenvectors of the covariance matrix.
[0077] 5) Arrange the eigenvectors from top to bottom row by row according to the size of the corresponding eigenvalues, and take the first b rows (12 in this embodiment) to form matrix P.
[0078] 6) Dimensionality reduction of matrix X is performed based on matrix P. In this embodiment, Y = PX is used to reduce the dimension of matrix X to b dimensions, and each data point becomes h after dimension reduction. i * This forms the dimensionality-reduced feature variable dataset D. p .
[0079] 7) For the dimensionality-reduced feature variable dataset D p For each data point d, calculate the distance between data point d and other samples belonging to the same category (feature flag value is unsatisfactory) according to Euclidean algorithm, and find the b samples closest to data point d as the b nearest neighbors of data point d.
[0080] 8) Randomly select C samples from b neighboring samples, and calculate the data augmentation sample h using the following formula. new * .
[0081]
[0082] Among them, h j * The sample is selected from b neighboring samples.
[0083] 9) Using the inverse operation of principal component analysis (PCA), iPCA, to augment the data in sample h. new * Data d restored to its original dimensions aug .
[0084] 10) Divide the data d of each original dimension aug Put into augmented sample dataset D aug Then, augment the sample dataset D. aug Place it into the feature variable dataset D to obtain the enhanced dataset.
[0085] S3. Based on the enhanced dataset, use the Xgboost model to obtain new features; reorganize the new features with the feature variable dataset D, and input the reorganized data into the logistic regression classifier to obtain the prediction results of customer business satisfaction.
[0086] This step utilizes the XgBoost model to perform feature transformation and generate new features. Finally, the new features are reconstructed with the original features and used as input features for a logistic regression model to predict customer satisfaction. Specifically, this includes:
[0087] 1) Set the XGBoost model parameters, including the number of iterations and the training step size. `n_estimators` controls the number of iterations during model training; too many or too few iterations will lead to overfitting. `learning_rate` is the training step size, which should also be set appropriately; too large or too small a step size is not conducive to finding the optimal fit point. In this example, `n_estimators` is set to 100 and `learning_rate` to 0.05.
[0088] 2) The enhanced dataset is split into training and test sets in a 3:1 ratio. The training set is used as the input to the Xgboost model. Finally, the feature vectors generated by the Xgboost model are processed by one-hot encoding to generate new features.
[0089] 3) Recombine the new features generated in the previous step with the original features (i.e., the feature variable dataset D), and input the recombined features into the Logistic Regression classifier. The output of the Logistic Regression classifier is the prediction result of customer business satisfaction.
[0090] This step can serve as an early warning system for predicting customer business satisfaction, and can improve the quality of services in the service area.
[0091] S4. Establish customer behavior profiles and conduct cluster analysis of customers from the perspective of value stratification to achieve precise marketing for power supply units, thereby promoting the improvement of service quality.
[0092] This step extracts customer payment data, constructs the RFM index and K-means clustering model, first calculates the three-dimensional RFM index, and uses it as input to the K-means model for clustering to stratify customers by value. The specific steps are as follows:
[0093] 1) Construct three indicators, R (Recency), F (Frequency), and M (Money), using customer payment data samples. Construct a dataset based on the RFM indicator. Data samples in the constructed dataset are shown in Table 3.
[0094] Table 3
[0095]
[0096] R (Recency): The time interval between the user's most recent payment, in days, calculated by subtracting the date of the most recent payment from the date of the statistical period;
[0097] F (Frequency): The user's cumulative payment frequency over a recent period. This example uses data from the past year.
[0098] M (Money): The user's cumulative payment amount within a recent period.
[0099] 2) Using the dataset constructed based on the RFM metric as input, a K-means clustering model is built, and the silhouette coefficient method is used iteratively to select a suitable number of clusters K (in this embodiment, the number of clusters K is chosen to be 3). The specific process is as follows:
[0100] Assume the data to be classified has been clustered using a clustering algorithm, resulting in K clusters. For each sample point i in each cluster, calculate its silhouette coefficient. Specifically, the following two metrics need to be calculated for each sample point i:
[0101] a(i): The average distance from sample point i to other sample points belonging to the same cluster. The smaller a(i) is, the greater the probability that sample point i belongs to that cluster.
[0102] b(i): The minimum average distance bij from sample point i to all samples in other clusters Cj, i.e.:
[0103] b(i) = min(bi1,bi2,…,bik)
[0104] The contour coefficient of sample point i is:
[0105] s(i)=b(i)-a(i) / max(a(i),b(i))
[0106] The average silhouette coefficient of all sample points i is the total silhouette coefficient S of the clustering result. S∈[-1,1], and the closer the silhouette coefficient S is to 1, the better the clustering effect.
[0107] 3) Randomly select N data objects as the initial cluster centers;
[0108] 4) Calculate the distance of each of the remaining data objects to the N initial cluster centers, and assign each data object to the cluster of the nearest cluster center;
[0109] 5) Adjust the new class and recalculate the center of the new class.
[0110] Repeat steps 3) and 4) to see if the cluster centers converge (remain unchanged). If they converge or the number of iterations is reached, stop the loop.
[0111] S5. Calculate the values of each indicator used for comprehensive evaluation in each service, then construct an entropy algorithm to obtain the weight of each indicator, and thus calculate the service quality score of each station area.
[0112] Based on the importance of each business, the smaller the entropy of a single business indicator, the greater the degree of variation in the indicator value, the greater the amount of information provided, and the greater its role in the comprehensive evaluation. Therefore, the weight of this indicator should also be greater. A total of 171 indicators in 7 categories, including power supply quality in distribution areas, 95598 work orders, and business processing satisfaction, were analyzed. The main indicators include: data collection coverage, data collection success rate, line loss rate, theoretical line loss, electricity meter ratio, power outage duration, number of power outage records, power outage duration, number of power outage records, and voltage anomaly duration. Sample data is shown in Table 4.
[0113] Table 4
[0114]
[0115] In this embodiment, the specific steps are as follows:
[0116] 1) Based on the principle of entropy method, establish a mathematical model for the index system, specifically as follows:
[0117]
[0118] In the formula, X nm This indicates the specific numerical value of the indicator.
[0119] 2) Based on the indicators, select positive and negative values for dimensionless data processing, where the positive indicators are:
[0120]
[0121] Negative indicators are:
[0122]
[0123] 3) Calculate the weight of a specific scheme under the indicator. The specific calculation method is as follows:
[0124]
[0125] 4) Calculate the information quotient and redundancy of the indicators. The calculation method is as follows:
[0126]
[0127]
[0128] 5) Calculate the indicator weights, the calculation method is as follows:
[0129]
[0130] The output weight values in this embodiment are shown in Table 5:
[0131] Table 5
[0132]
[0133] 6) The service quality score for each station area is obtained by weighting and summing the above indicator weights and the weight values after standardization.
[0134] S6. Based on the prediction results of customer business satisfaction, the cluster analysis results of customers, and the service quality scores of each service area, adjust the service quality of the corresponding service area.
[0135] Example 2
[0136] Based on the same inventive concept as Embodiment 1, this embodiment provides a machine learning-based service quality management system for distribution centers, including the following modules:
[0137] The dataset acquisition module is used to analyze and process power work order data to obtain the feature variable dataset D;
[0138] The data augmentation module is used to construct a data simulation neighborhood interpolation algorithm based on the principle of minimizing neighborhood risk to achieve data augmentation and obtain an augmented dataset.
[0139] The combined model prediction module is used to obtain new features from the enhanced dataset using the Xgboost model; the new features are then recombined with the feature variable dataset, and the recombined data is input into the logistic regression classifier to obtain the prediction results of customer business satisfaction.
[0140] The customer clustering analysis module is used to build customer behavior profiles and perform clustering analysis on customers from the perspective of value stratification.
[0141] The service quality score calculation module is used to calculate the values of various indicators used for comprehensive evaluation in various services, construct an entropy algorithm to obtain the weight of each indicator, and calculate the service quality score of each station area.
[0142] The integrated module is used to combine the predicted results of customer business satisfaction, the results of customer cluster analysis, and the service quality scores of each service area to adjust the service quality of the corresponding service area.
[0143] The data augmentation process includes:
[0144] Form matrix X from the n d-dimensional data points of the feature variable dataset D;
[0145] Perform zero-mean processing on each column of matrix X;
[0146] Find the covariance matrix of matrix X;
[0147] Find the eigenvalues and corresponding eigenvectors of the covariance matrix;
[0148] Arrange the eigenvectors from top to bottom according to the size of their corresponding eigenvalues, and take the first b rows to form matrix P;
[0149] Dimensionality reduction of matrix X is performed based on matrix P, reducing matrix X to b dimensions, resulting in the dimensionality-reduced feature variable dataset D. p ;
[0150] For the dimensionality-reduced feature variable dataset D p For each data point d, calculate the distance between other samples belonging to the same category and data point d, and find the b samples that are closest to data point d, which are then considered as the b nearest neighbors of data point d.
[0151] Randomly select C samples from b neighboring samples and calculate the data augmentation samples;
[0152] Using iPCA, the inverse operation of principal component analysis (PCA), the data augmented samples are restored to their original dimensions.
[0153] Put the data from each original dimension into the augmented sample dataset, and then put the augmented sample dataset into the feature variable dataset D to obtain the augmented dataset.
[0154] In this embodiment, the prediction process of the combined model prediction module includes:
[0155] Configure XGBoost model parameters, including the number of iterations and the training step size;
[0156] The enhanced dataset is split into a training set and a test set. The training set is used as the input to the Xgboost model. The feature vectors generated by the Xgboost model are processed by one-hot encoding to generate new features.
[0157] The new features are recombined with the feature variable dataset, and then input into a logistic regression classifier. The output of the logistic regression classifier is the prediction result of customer business satisfaction.
[0158] The modules described above in this embodiment are used to implement the steps of embodiment 1, and their detailed implementation process can be found in embodiment 1, which will not be repeated here.
[0159] The above embodiments are preferred embodiments of the present invention, but the embodiments of the present invention are not limited to the above embodiments. Any changes, modifications, substitutions, combinations, or simplifications made without departing from the spirit and principle of the present invention shall be considered equivalent substitutions and shall be included within the protection scope of the present invention.
Claims
1. A machine learning-based method for managing service quality in distribution areas, characterized in that, Includes the following steps: The power work order data was analyzed and processed to obtain the feature variable dataset D; Based on the principle of minimizing neighborhood risk, a data simulation neighborhood interpolation algorithm is constructed to achieve data augmentation and obtain an augmented dataset. Based on the enhanced dataset, new features are obtained using the Xgboost model; The new features are recombined with the feature variable dataset, and the recombined data is input into a logistic regression classifier to obtain the prediction results of customer business satisfaction. Establish customer behavior profiles and perform cluster analysis on customers from the perspective of value segmentation; Calculate the values of each indicator used for comprehensive evaluation in each business segment, construct an entropy algorithm to obtain the weight of each indicator, and calculate the service quality score of each distribution area. Based on the combined prediction results of customer business satisfaction, the results of customer cluster analysis, and the service quality scores of each service area, the service quality of the corresponding service area is adjusted. The data augmentation process includes: Form matrix X from the n d-dimensional data points of the feature variable dataset D; Perform zero-mean processing on each column of matrix X; Find the covariance matrix of matrix X; Find the eigenvalues and corresponding eigenvectors of the covariance matrix; Arrange the eigenvectors from top to bottom according to the size of their corresponding eigenvalues, and take the first b rows to form matrix P; Dimensionality reduction of matrix X is performed based on matrix P, reducing matrix X to b dimensions, resulting in the dimensionality-reduced feature variable dataset D. p ; For the dimensionality-reduced feature variable dataset D p For each data point d, calculate the distance between other samples belonging to the same category and data point d, and find the b samples that are closest to data point d, which are then considered as the b nearest neighbors of data point d. Randomly select C samples from b neighboring samples and calculate the data augmentation samples; Using iPCA, the inverse operation of principal component analysis (PCA), the data augmented samples are restored to their original dimensions. Put the data from each original dimension into the augmented sample dataset, and then put the augmented sample dataset into the feature variable dataset D to obtain the augmented dataset.
2. The method according to claim 1, characterized in that, Based on the enhanced dataset, new features are obtained using the Xgboost model; The new features are recombined with the feature variable dataset, and the recombined data is input into a logistic regression classifier to obtain predictions of customer business satisfaction; including: Configure XGBoost model parameters, including the number of iterations and the training step size; The enhanced dataset is split into a training set and a test set. The training set is used as the input to the Xgboost model. The feature vectors generated by the Xgboost model are processed by one-hot encoding to generate new features. The new features are recombined with the feature variable dataset, and then input into a logistic regression classifier. The output of the logistic regression classifier is the prediction result of customer business satisfaction.
3. The method according to claim 1, characterized in that, Establish customer behavior profiles and perform cluster analysis on customers from a value segmentation perspective, including: The RFM indicator was constructed using a sample of customer payment data, and a dataset was built based on the RFM indicator. Using a dataset constructed based on the RFM index as input, a K-means clustering model is built, and the silhouette coefficient method is used iteratively to select an appropriate cluster K. Randomly select N data objects as the initial cluster centers; Calculate the distance of each remaining data object to the N initial cluster centers, and assign the data object to the cluster class containing the nearest cluster center; Adjust the new class and recalculate the center of the new class.
4. The method according to claim 1, characterized in that, The indicators used for comprehensive evaluation include: data collection coverage, data collection success rate, line loss rate, theoretical line loss, electricity meter ratio, power outage duration, number of power outage records, power outage duration, number of power outage records, and voltage anomaly duration.
5. The method according to claim 4, characterized in that, Calculate the values of each indicator used for comprehensive evaluation in various business operations, construct an entropy algorithm to derive the weight of each indicator, and calculate the service quality score for each distribution area; including: Based on the principle of entropy method, a mathematical model for the index system is established; Based on the indicators, positive and negative values are selected for dimensionless data processing; Calculate the weight of a particular scheme relative to that indicator; Calculate the information quotient and redundancy of the indicators; Calculate the indicator weights; The service quality score for each district is obtained by weighting and summing the indicator weights and the standardized weight values.
6. The method according to claim 1, characterized in that, The data sources for power work orders include complaint work order data from provincial power companies and call work order data from the State Grid customer service center.
7. A machine learning-based service quality management system for distribution areas, characterized in that, Includes the following modules: The dataset acquisition module is used to analyze and process power work order data to obtain the feature variable dataset D; The data augmentation module is used to construct a data simulation neighborhood interpolation algorithm based on the principle of minimizing neighborhood risk to achieve data augmentation and obtain an augmented dataset. The combined model prediction module is used to obtain new features based on the enhanced dataset using the Xgboost model; The new features are recombined with the feature variable dataset, and the recombined data is input into a logistic regression classifier to obtain the prediction results of customer business satisfaction. The customer clustering analysis module is used to build customer behavior profiles and perform clustering analysis on customers from the perspective of value stratification. The service quality score calculation module is used to calculate the values of various indicators used for comprehensive evaluation in various services, construct an entropy algorithm to obtain the weight of each indicator, and calculate the service quality score of each station area. The integrated module is used to combine the predicted results of customer business satisfaction, the results of customer cluster analysis, and the service quality scores of each service area to adjust the service quality of the corresponding service area. The data augmentation process includes: Form matrix X from the n d-dimensional data points of the feature variable dataset D; Perform zero-mean processing on each column of matrix X; Find the covariance matrix of matrix X; Find the eigenvalues and corresponding eigenvectors of the covariance matrix; Arrange the eigenvectors from top to bottom according to the size of their corresponding eigenvalues, and take the first b rows to form matrix P; Dimensionality reduction of matrix X is performed based on matrix P, reducing matrix X to b dimensions, resulting in the dimensionality-reduced feature variable dataset D. p ; For the dimensionality-reduced feature variable dataset D p For each data point d, calculate the distance between other samples belonging to the same category and data point d, and find the b samples that are closest to data point d, which are then considered as the b nearest neighbors of data point d. Randomly select C samples from b neighboring samples and calculate the data augmentation samples; Using iPCA, the inverse operation of principal component analysis (PCA), the data augmented samples are restored to their original dimensions. Put the data from each original dimension into the augmented sample dataset, and then put the augmented sample dataset into the feature variable dataset D to obtain the augmented dataset.
8. The system according to claim 7, characterized in that, The prediction process of the combined model prediction module includes: Configure XGBoost model parameters, including the number of iterations and the training step size; The enhanced dataset is split into a training set and a test set. The training set is used as the input to the Xgboost model. The feature vectors generated by the Xgboost model are processed by one-hot encoding to generate new features. The new features are recombined with the feature variable dataset, and then input into a logistic regression classifier. The output of the logistic regression classifier is the prediction result of customer business satisfaction.